阿里发布 Qwen3 系列

2025 年初的 DeepSeek-R1 制造了一场全球性震荡：它证明了推理能力可以靠开源模型实现，还顺带让「中国开源」成为世界焦点。但 R1 之后，一个问题悬在行业上空：开源模型能把推理做到多高？阿里通义团队在 2025 年 4 月底给出了自己的回答。

4 月 29 日，Qwen3 系列发布。这个家族从 0.6B 一直排到 235B——后者采用 MoE 架构，是当时最大的开源模型之一。更重要的是，Qwen3 原生支持思考和非思考两种模式：需要严谨推理时开启思考，需要快速响应时关掉它。工具调用、长上下文一应俱全，而且全部权重开源。

Qwen3-235B 在多项评测上登顶开源模型榜单，让「开源旗舰」四个字的分量又重了一分。但它的意义不止于性能数字。它把 DeepSeek-R1 开创的「开源推理」变成了一条成熟的产品线，用一套覆盖全尺寸的家族，向世界展示：中国开源模型不但能做旗舰，还能把推理做成标配。

对开发者来说，Qwen3 意味着「旗舰级推理模型免费拿」——全球团队都能在 235B 的基座上做微调、做部署、做创新。闭源模型剩下的优势，在一次次这样的发布中被不断压缩。定价压力也随之传导到全行业，谁还在用「高不可攀」作为卖点，谁就会显得落伍。

回看 Qwen3，它是中国开源模型登顶之路上的一个里程碑。从 Qwen2 的「系列化」，到 Qwen3 的「旗舰+推理」，阿里用两次发布完成了从追赶者到领跑者的转身。当世界后来讨论「开源到底能不能追平闭源」时，Qwen3 已经用行动给出了越来越明确的答案。

DeepSeek-R1 in early 2025 created a global shock: it proved reasoning capability could be achieved with open models and, along the way, made "Chinese open source" a world focus. But after R1, a question hung over the industry: how high can open models push reasoning? Alibaba's Qwen team gave its answer in late April 2025.

On April 29, the Qwen3 family was released. It ranged from 0.6B all the way up to 235B—the latter using a MoE architecture and being among the largest open models of its time. More importantly, Qwen3 natively supported both thinking and non-thinking modes: enable thinking for rigorous reasoning, disable it for fast response. Tool use and long context were all included, and all weights were open.

Qwen3-235B topped open-model leaderboards on many benchmarks, adding weight to the phrase "open-source flagship." But its significance went beyond the numbers. It turned the "open reasoning" pioneered by DeepSeek-R1 into a mature product line, showing the world with a full-size family that Chinese open models could not only build flagships but make reasoning standard equipment.

For developers, Qwen3 meant "flagship-grade reasoning model for free"—teams worldwide could fine-tune, deploy, and innovate on the 235B base. The remaining advantages of closed models were compressed again with each such release. Pricing pressure also spread industry-wide; anyone still using "unreachable" as a selling point looked increasingly out of date.

Looking back, Qwen3 is a milestone on Chinese open models' road to the top. From Qwen2's "serialization" to Qwen3's "flagship plus reasoning," Alibaba completed the turn from chaser to leader in two releases. When the world later debated "can open source actually catch up with closed source," Qwen3 had already answered with increasingly clear action.